Both rows that could not go red are decided, each by measurement.
test_lexicon.py::test_redos_pathological_subagent_input_returns_fast is REVIVED,
not retired. The row was not dead because the seed form is safe -- it was dead
because both earlier payloads made the prefix match at ONE start position, and
the cost is per-prefix-match. Repeating `spawn an agent that ` instead makes it
match K times, each driving its own O(N) lazy scan for a keyword never supplied:
K x O(N) against the seed's `(?:.*?\s+)?`, K x O(1) against the shipped
`{0,12}?` bound. Measured through scan_lexicon at 1500/3000/6000/12000 words:
seed 0.091/0.283/1.085/4.091s (exponent 1.92), shipped 0.047/0.051/0.094/0.190s
(exponent 1.01). Verified red with the seed form patched in: 4.21s against the
2.0s bound. The nesting the old comment blamed was a red herring.
test_output.py::test_pathological_input_returns_within_a_bound moves to CPU time
with a 20.0s bound, and the "or a hang" half of its claim is retired. The wall
clock was kept because a blocking hang burns no CPU -- true in general, and
inapplicable to a path with no open(), socket, subprocess, thread, lock or sleep
anywhere on it. Same payload, idle vs ~4x oversubscription: wall 3.30 -> 21.63s
(2x over the old 10.0s bound), cpu 3.30 -> 7.62s. It guarded a mode it could not
have while paying the full false-red premium. No in-repo vulnerable form can
turn this row red, so the bound was proved live against what it actually guards
-- a future pattern quadratic on long runs, `A+\s*EXFILTRATE` -- which failed it
at 64.77s CPU, 3.2x over.
redos_clock.py and the clock's pin test both documented this row as the
deliberate wall-clock holdout; both corrected.
792 passed, 6/6 documented gaps hold.
`5667063` moved test_output.py's ReDoS bounds off the wall clock, because a
loaded machine steals wall seconds without adding any cycles and two rows
failed at 2.24s / 3.66s against a 2.0s bound while census had the CPU. The
remaining ten bounds in five other files still ran on `time.monotonic()` and
carried the same defect. They now share ONE clock.
The clock is IMPORTED, not copied: `tests/redos_clock.py`. Five private copies
would leave four of them unpinned -- the instrument test
(test_the_redos_clock_ignores_time_this_process_did_not_spend) can only pin the
implementation it calls, and the suite already holds that rule for the code it
measures.
Every ported row was verified the only way a time bound can be: the vulnerable
form patched back in, red demanded, `git checkout --` after. Measured against
the 2.0s bound (3.0s for the url arm):
active_content long-attr `{0,63}` -> `*` RED
neutralize long-attr same patch RED
output gate long-attr same patch 12.41s
okf link graph `[^\]\[]` -> `[^\]]` 6.91s
sanitize comment str.find -> `<!--.*?-->` 17.56s
lexicon md-link-anchor-text 319.14s
lexicon md-link-anchor-url 8.55s
lexicon md-link-ref-comment 37.82s
Two rows did not go red, for two different reasons.
test_sanitize.py::test_legitimate_comment_heavy_document is the legitimate SIDE
of a separation, not a second pin on the defect: closed comments never withhold
the required literal, so the lazy form runs it in 0.016s. Recorded in place.
test_lexicon.py::test_redos_pathological_subagent_input_returns_fast is DEAD --
the same zero-signal shape the `<a ` carrier had, found by the same method. The
seed form is `(?:.*?\s+)?` (llm-security 7.8.0, injection-patterns.mjs:84) and
this repo has never carried it: the bounded `{0,12}?` port is in the pattern
table's first commit. Patched in by hand at the row's own size: shipped 0.135s
vs seed 0.113s, separation 1.2x. Not the keyword gate either -- a variant that
reaches the inner branch stays linear over four doublings (exponent ~1.0),
because the nesting is one lazy run inside an OPTIONAL group, never a repeated
one. Left standing with the measurement written into it; picking a new carrier
is an operator call, like the wall-clock row above it.
The dead sibling row named in STATE is fixed: test_active_content.py's
long-attribute row swaps carrier `<a ` -> `<script `, for the reason `5667063`
established on its composed-gate twin -- 0.7.0's own no-URL narrowing put `<a>`
in `_URL_AFFORDANCE_TAGS`, so the tag returns inert BEFORE its body reaches the
arm the row guards. Re-measured here, not inherited: `<a ` 0.041s and NO
findings against the vulnerable form; `<script ` 19.349s against 0.052s
shipped, 373x apart.
`test_pathological_input_returns_within_a_bound` deliberately keeps its wall
clock (operator decision): it claims to catch a hang, and only a wall clock
catches one.
792 tests, 129/129, 6/6.
The input-path duty `8deca93` scoped. All 83 lexicon patterns measured arm by
arm; two are quadratic, same shape 0.3.2 fixed -- a run in front of a required
literal that may cross the pattern's own opening anchor. Exponent 1.98 over five
points, so quadratic, not exponential.
markdown:link-anchor-injection `[` 1.91s @8k ~8.3h at the cap
markdown:link-anchor-injection `[system](` 0.006s @8k ~89s at the cap
markdown:link-ref-comment `[//]: # (` 0.22s @8k ~1.0h at the cap
Not input-path-only: `scan_lexicon` runs on the output path, so `scan_output("["
* 100_000)` took 334.7s. 0.3.2's "last quadratic site on the output path" was
false when written -- its sweep drove `[` only through `scan_active_content`.
Fix is anchor exclusion, not bounding (bounding attacker-controlled content is a
one-line bypass). The excluded char is `(`, not the obvious `[`: excluding `[`
drops `[//]: # (see [x] then ignore this)`, which no other pattern catches. The
anchors contain `(` too, so it telescopes at zero measured recall cost.
N is per row deliberately. The URL arm ran 0.9s UNFIXED at N=100_000 -- under the
2.0s bound, so that row could not have failed. Measured at N=300_000 instead,
where crafted (8.10s) and legitimate (0.926s) separate 8.8x.
The output gate claimed LLM10 self-safety on the grounds that its patterns have
no nested quantifiers. True, and irrelevant: nesting is not what makes these
blow up. A run in front of a REQUIRED literal, reachable from a short anchor, is
enough -- crafted input repeats the anchor and never supplies the literal, so
every start position rescans the tail. Quadratic, not exponential, and the
max_scan_chars cap does not help: it bounds the input, and quadratic work on a
bounded input is still hours.
Measured, not argued. `<a:` x 100_000 took 23.4s in AUTOLINK_RE alone; the
composed gate on that payload took 458.7s, extrapolating to ~5.7 hours at the
1_000_000-char input the gate itself accepts. Size-matched ordinary prose runs
0.31s, so the separation is 18x-660x -- unlike the blob in the neighbouring
test, which is the *faster* side of prose and never exercised backtracking.
Two fixes, chosen per pattern rather than uniformly:
- active_content + lexicon JSON (15 runs): exclude the character that opens the
pattern's own anchor (`[` for markdown, `<` for tags), so a run cannot reach
past the next start position and the per-start costs telescope. Verified to
cost no recall: long URLs, long alt text, and `<` inside a quoted attribute
all still match. Bounding instead would have been linear too but wrong here --
the content is attacker-controlled, so padding past a bound would be a
one-line bypass of the EchoLeak class this table exists to catch.
- connstr egress (4 runs): bound the password at MAX_CONNSTR_VALUE. The
exclusion fix is unavailable -- the anchor character is `/` and passwords
containing `/` are the common case (measured: they match today). The residual
miss is a credential over 256 chars; a token that long is still caught by
egress:jwt-token.
hybrid-xss:script-tag had neither option: its run is the script BODY, which may
legitimately contain `<`. It now matches the opening tag and drops the
`</script>` requirement. That also closes a fail-open -- `<script>alert(1)`
unclosed was silently missed -- at the cost of flagging prose that merely
mentions `<script>`, now documented.
Found by the composed-gate test staying red after every individual scanner was
already linear: the lexicon's six html-obfuscation patterns were the remaining
813x. A per-scanner test alone would have shipped that.
662 passed (was 642), and faster than before the fix.
Build order step 4 — the load-bearing port from the llm-security seed
(injection-patterns.mjs + string-utils.mjs), stdlib-only.
- injection_lexicon.json: 83 patterns (CRITICAL/HIGH/HYBRID/MEDIUM) as the
single source of truth (regex + id + severity + owasp + desc), compiled once
by a thin loader. Decoupled from the engine for a future TS port.
- scan_lexicon(text, source, max_scan_chars) -> Report: matches every pattern
against a deduped variant set (raw / normalized / homoglyph-folded / rot13),
plus unicode-tag presence signal and the cognitive-load trap.
- normalize_for_scan chain ported: unicode-tags -> bidi -> HTML-entities ->
unicode/hex/URL escapes -> whole-string base64 (reuses entropy.try_decode_base64)
-> collapse letter-spacing; plus fold_homoglyphs / rot13.
- Self-safety (OWASP LLM10): input-size cap (scan prefix + flag oversize) and
ReDoS-safe port — the two nested-.*? sub-agent patterns bounded to
(?:\S+\s+){0,N}?; verified true positives still fire.
- Non-Latin data (homoglyph map, BIDI block) built from explicit code points;
JSON non-ASCII kept as \uXXXX escapes.
24 tests; 55 green total.
[skip-docs]: README positioning + honest-limitations is a deliberate build-order
step-11 deliverable (steps 1-3 likewise left README frozen). README status line
("pre-implementation") is stale and flagged for the step-11 refresh.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K8GmKRCdsPjWYAKWsNgeQS
